Parallel iterative solution methods for linear finite element computations on the CRAY T3D

摘要

This paper describes the implementation of a finite element program on the massively parallel CRAY T3D. We have used a domain decomposition approach to parallelize the computations. The paralleliza-tion of the discretization is straightforward. The solution of the resulting system of equations is more difficult to parallelize. For this we use iterative techniques. Two communication operations have to be performed for these methods: global communication for the inner product and local communication for the matrix-vector multiplication. We discuss the routines for communicating data to other processors and give timing results for these routines. Experiments show satisfactory speed-ups if the number of processors is increased, both for the discretization phase and for the solution phase.